DeepSeek plans to pretrain V5 on Huawei Ascend 950 chips
Prepared by Sintelligent with the help of SI. Published in SI: . Updated: .
If the plan is confirmed and works, Chinese labs will be able to complete the full training cycle for frontier models without Nvidia rather than only running finished models on domestic chips.
According to Caiwen News and STAR Market Daily, citing industry sources, V5 could become DeepSeek's first large model pretrained from scratch on Ascend; Huawei's platform was previously used mainly for inference, while a separate research consortium has already performed full-parameter post-training of V4-Pro on Ascend 910C. The sources expect the project to materialize in 2027 and link it to restricted access to Nvidia's latest chips. The plan is at an early stage and has not been confirmed by the companies; its technical outcome will depend on cluster stability, software-stack maturity, and the resulting model's capabilities.
Sources 2